Tencent Hunyuan ships Hy Realtime, a voice model that decides when to search

Tencent Hunyuan has released Hy Realtime, a new realtime voice model. According to Pandaily, the model decides for itself when a conversation calls for an outside search, and it can sing on request.

Both features point at the same shift: a voice model is no longer just a reader of text. Letting the model decide whether to search mid-conversation means it has to weigh answering directly against looking something up first, instead of triggering a lookup on every turn and paying the latency cost.

Singing on request suggests the generation side keeps control over pitch and prosody well beyond plain narration. It reads as a demo flourish, but it also shows how finely the model can steer audio detail.

Realtime speech is becoming a distinct front in the model race. Text chat is judged on answer quality, while voice adds latency, interruption handling and turn-taking, where a single stall breaks the illusion of conversation. Handing the search decision to the model is one way to make that loop feel less mechanical.

Shipping Hy Realtime as a standalone release also suggests Tencent treats the voice entry point as its own product surface rather than an accessory to a text model. Public material around the launch still centers on capability demos, with no details yet on model size, supported languages, access terms or pricing.

Three things are worth watching: whether Hy Realtime is opened up through an API, how quickly it shows up inside Tencent's own products, and how fast rivals answer with their own mixes of self-directed search and expressive audio.